Graph-Based Real-World Event Extraction In Social Stream For Disaster Management

Hazard analysis plays an important role in disaster management. There is a great need to continuously monitor the situations of potential hazard activities as well as their corresponding society activities for emergency decision. Due to the fact that cyberspace and real world affect and influence each other, user-contributed messages on social media sites can be used as sensor data for real world event or activity perception. In this paper, we focus on context-aware event extraction from microblogs. Compared with traditional methods using vector-based context representation to filter and group noisy and scattered messages for event extraction, we propose a Hazard-Centroid Context Graph (HCCG) based method, where both word-level and stream-level contexts of concerned hazards can be effectively modeled and exploited to form an informative and straightforward graphical representation of real-world event. The proposed method is tested on social streams about Ebola in Sina Weibo, the Chinese counterpart of Twitter. Experimental results demonstrated the effectiveness of our proposed method.Based on the proposed approach, this paper adopts a two stage framework-clustering first and then classification to solve the event recognition problem. The first stage clustering similar texts together and the second stage classify real-world event clusters from the non-event clusters. In consideration of the real-time nature of social media, we deploy our system in Streaming computing platform for better responding performance.To utilize the social media information better, we also integrate sentiment analysis in our system to distinguish the subjective detected events from the clustering stream with quantization on negative and positive emotion so that we can have a more direct view of the public emotion trend.